2015年9月18日開催 GTC Japan 2015 講演資料
エヌビディア合同会社
エンタープライズプロダクト事業部 シニアソリューションアーキテクト Jeremy Main
A walk through of the techniques to monitor existing workstation workloads to create data-driven estimates of recommended user density levels based on the GPU requirements, frame buffer utilization and other factors as well as methods to confirm GPU resource utilization to ensure excellent performing NVIDIA GRID vGPU enabled virtual machines.
2. First Considerations
Understand the existing non-VDI environment and workloads
Workstation model, CPU generation, CPU speed, memory, storage
GPU(s) used, how it was selected or upgraded
Applications used, number of displays, special input devices
Capture system and GPU performance data and review with user
Provide a data-driven recommendation to gain acceptance
Segment users GPU requirements and size appropriately
3. First Considerations
Server-rendered FPS vs. remotely delivered FPS
Define the per-user network bandwidth requirements
Explain the impact of 30 FPS vs. 60 FPS delivered framerates
Start with a less dense vGPU profile and access user performance
Increase density until performance does not meet acceptance tests
Require acceptance testing with real workloads not benchmarks
Log and monitor GPU utilization on host
Agree on metrics, don’t use subjective criteria…
4. Understanding your Applications
CPU, memory and storage requirements
GPU rendering and frame buffer requirements
“perfmon” on existing workstation or GPU pass-through VM
Memory -> Available MBytes
Processor -> % Processor Time
NVIDIA_GPU
% GPU Usage
Total Memory (MB)
Available Memory (MB)
5. Catia V5-6R2012 K5000
Try to limit the number of slides you use
Keep text to a minimum
Instead, speak more to your audience
(engage them with anecdotes/enthusiasm/eye contact)
Try not to read your points verbatim; bullet points should
be used for key points only
Use images/graphics to help convey your message
12. GPU Utilization
0
50
100
% CPU
% GPU
Benchmarks
Push GPU capabilities to their limits
and may be heavily dependent on
single thread performance
Applications
Modeling operations and view
manipulation consist of periods of
activity and inactivity
0
50
100
% CPU
% GPU
13. Monitoring vGPU
Within the VM “NVIDIA_GPU” counter, “% GPU Usage” is not supported
Monitor at hypervisor-level using nvidia-smi command
$ nvidia-smi --query-gpu=¥
timestamp,name,pci.bus_id,driver_version,pstate,¥
pcie.link.gen.max,pcie.link.gen.current,temperature.gpu,¥
utilization.gpu,utilization.memory,memory.total,¥
memory.free,memory.used --format=csv -l 5
Prepend with “timeout –t” with the number of seconds to run
14. Monitoring protocol HDX3D-Pro FPS
wmic /NameSpace:¥¥root¥citrix¥hdx Path Citrix_VirtualChannel_Thinwire_Enum Get
Component_Fps /every:5
Verify consistent frame
rate delivery based on
XenDesktop policy
parameters